Comparison of amyloid PET measured in Centiloid units with neuropathological findings in Alzheimer’s disease

Current standard-of-truth (SoT) diagnosis of Alzheimer’s disease (AD) largely depends on neuropathological demonstration of brain amyloid-beta (Aβ) plaques and tau tangles [1]. Recently, research criteria for detection of AD have emphasised the importance of amyloid and tau positron emission tomography (PET) imaging biomarkers [2, 3].

¹¹C-Pittsburgh Compound B (¹¹C-PiB), ¹⁸F-florbetaben (FBB), ¹⁸F-florbetapir, ¹⁸F-flutemetamol and ¹⁸F-NAV4694 (NAV) are PET tracers that demonstrate binding to brain Aβ in AD from the preclinical AD stage onwards, with good sensitivity and specificity as biomarkers of antemortem AD pathology and predictors of progression to AD dementia [4‐10]. Amyloid PET scans are used for inclusion and monitoring in AD-modifying clinical therapy trials and to aid clinical diagnosis and prognostication [11, 12].

Variability in tracers, PET scanners, procedural factors, and analysis methods across imaging centres have driven attempts for quantitative standardisation of amyloid PET results. Klunk and colleagues [13] derived a scale of “Centiloid” units (CL) for standardised reporting of amyloid imaging. CL values range beyond the “anchor-point” of 0, representing young healthy controls, and 100, representing the amyloid burden present in average mild to moderate severity dementia due to AD. This important work allows for amyloid PET scans across different sites to yield comparable data.

Comparison of neuropathological data with positive or negative amyloid PET scans based on expert visual read has been performed for ¹¹C-PiB, ¹⁸F-florbetapir, ¹⁸F-florbetaben and ¹⁸F-flutemetamol [7, 9, 14‐18]. In vivo biomarkers such as cerebrospinal fluid Aβ, tau PET, and volumetric MRI have been compared with amyloid PET in CL [19, 20]. Three recent studies have examined the performance of amyloid PET CL thresholds compared with SoT neuropathology. These studies reported thresholds for detection of moderate or frequent neuritic plaque ranging from 12 to 24 CL but did not correct for time elapsed between amyloid scan and death and had relatively few cases with CL values close to the threshold values [21‐23]. Only one compared to Alzheimer’s disease neuropathologic change (ADNC) rating [1] or to expert visual read report of a positive scan [21].

We aim to further define the accuracy of CL values when compared with SoT post-mortem neuropathological data on neuritic amyloid plaque density, ADNC rating, final clinicopathological diagnosis of AD, and visual reading threshold for a positive amyloid PET scan.

Improving accuracy of the detection of brain amyloid plaques is important for clinical trial enrichment in the quest to develop disease-modifying or curative treatment for the growing burden of Alzheimer’s disease. Contemporary clinical application can also better assist diagnosis, prognostication and planning for patients with cognitive disorders.

We have demonstrated that a threshold of 20.1 CL was optimal for the detection of “high” levels of neuritic plaque density, as determined by C score moderate or frequent classification. In other words, values of 20.1 CL or lower accurately reflected the absence of moderate or frequent plaques, with a high AUC. This threshold was not significantly altered when corrected for time between scan and post-mortem. CL results below this threshold should provide reassurance that patients are unlikely to have Alzheimer’s disease. This is reasonably concordant with the findings of Navitsky and colleagues, who determined a threshold of 24.1 CL with florbetapir for CERAD amyloid plaque classification of moderate or frequent vs sparse or none, in 59 individuals [22]. This is also concordant with the Centiloid analysis by Dore and colleagues of a florbetaben phase III post-mortem study in 66 individuals, which yielded a threshold of 19 CL for the same categorisation [23]. These thresholds are higher than those identified by La Joie et al., who determined a threshold of 12.2 CL for separating none or sparse plaques from moderate or frequent plaques; but oddly they found the same threshold for separating any plaques from no plaques in 179 individuals scanned with ¹¹C-PiB PET [21]. Those authors did, however, suggest 24 CL to be a more appropriate threshold for “identifying clinically meaningful Aβ burden” based on ADNC neuropathological criteria and Thal scores. These thresholds may be useful as cut-offs for clinical trial enrichment, or for guiding decision-making about commencing potential disease-modifying therapies when they become available.

For the detection of “any” amyloid plaque (sparse or more), the optimal threshold identified was 9.5 CL. Once again, this value only marginally increased when corrected for time between scan and post-mortem. This threshold is similar to two standard deviations of young controls as determined by Klunk and colleagues, equalling 8.68 CL [13] and 12.2 CL [21] for PiB and FBB respectively. This suggests there are no differences in amyloid tracer binding between young and old individuals when no amyloid plaques are present, indicating no substantial increase in non-specific binding of these tracers nor significant changes in tracer kinetics with normal ageing.

A threshold of 26 CL exactly matched expert visual read of positive vs negative scan. This is consistent with good concordance (97%) noted by Leuzy and colleagues [19] between visual read and PiB CL results. This visual threshold is slightly higher than the CL threshold for neuropathology indicating that some persons with significant AD pathology will have a visually negative scan and that scan quantification is of value when identifying persons with AD pathology for clinical trials.

Our expert visual reader results are comparable to those reported in several phase III trials of PET tracers that compared blinded expert visual reads of amyloid imaging with post-mortem data, where moderate to frequent neuritic plaques (a “high” classification in our study) were considered positive. Florbetaben visual reads were reported as having sensitivity of 97.9% and specificity of 88.9% [16]. Florbetapir visual reads yielded sensitivity of 92% and specificity of 100% within a 2-year window between imaging and autopsy [15]. Flutemetamol visual reads in one study correlated with both original and modified CERAD criteria, yielded respective sensitivities of 91.9% and 90.8%, and specificities of 87.0% and 90.0% [17]. Three expert readers in La Joie’s group [21] had 99% accuracy for scans above a threshold of 24 CL. When considering visual read, there are rare cases of clear focal amyloid uptake that may lead to a positive visual read but a low CL score but there were no such cases in our cohort.

In our cases with clinicopathological AD, a median CL value of 87.7 was found at post-mortem, but with significant variability as demonstrated by the IQR of ± 42.2 CL. Only one of these cases, scoring 20.1 CL (21.3 CL when corrected for time to death), returned a result under 45 CL, suggesting that a sensitivity cut-off for defining clinicopathological AD should be considerably higher than that for detecting “high” amyloid plaques alone. This agrees well with our finding of 49 CL as the optimal threshold for identifying cases that meet current neuropathological criteria for AD based on a comprehensive post-mortem brain examination, i.e. intermediate or high Alzheimer’s Disease Neuropathic Change (ADNC). Our exceptional case with only 21 CL had logopenic aphasia and 5.6 years between ¹¹C-PiB scan and death at which time frequent plaques were found. The PiB scan was reviewed and showed mild patchy cortical binding. The long interval between scan and post-mortem is note-worthy and raises the possibility of more rapid than usual plaque accumulation so that the correction for time elapsed from scan did not provide an accurate estimation of the plaque burden at the time of death. Our CL findings in clinicopathological AD compare to estimates obtained without pathological confirmation of AD. For instance, Leuzy et al. [19] demonstrated median PiB PET results of 47.5 CL for mild cognitive impairment and 84.1 CL for AD in their cohorts. We reviewed the 230 patients with a diagnosis of probable AD made by a clinical panel blinded to biomarker findings including amyloid PET in the Australian Imaging Biomarkers and Lifestyle study of ageing. The mean CL in these patients was 95 ± 30. Larger numbers in future studies would help confirm if a clinicopathological diagnosis of AD is indeed rare when under 45 CL.

The sex distribution in this study was predominantly male at 78%. No explanation for this is evident in our study, but a predominance of males was also reported in two of the three published CL neuropathological correlation studies [21, 22].

Our study adds further confirmation of these thresholds in a field where there are relatively low numbers of post-mortem results for correlation, especially for intermediate CL values. Additionally, we have adjusted for amyloid accumulation during the time between scan and post-mortem. Finally, only La Joie and colleagues [21] have also included a visual read comparison in a cohort with neuropathological assessment.

In our cohort, values < 9.5 CL accurately reflected the absence of any neuritic plaques, and > 20.1 CL indicated the presence of at least moderate plaque density. These neuropathology-based Centiloid thresholds may be used to exclude a diagnosis of AD and to define groups for early intervention and other disease specific trials. Approximately 50 CL or more best confirmed both neuropathological and clinicopathological diagnosis of Alzheimer’s disease.